Tactile World Model Enables Robot Error Correction

机器之心 · wechat · 2026-07-12

Synced introduced the tactile dexterous manipulation work TouchWorld by Yang Shuo's team at Poxiao Intelligence / Harbin Institute of Technology, Shenzhen. The core idea is to use tactile sensing simultaneously for "pre-action prediction" and "post-contact correction."

Predictive Tactile Goals: Before executing a task, the model predicts not only future visual states but also the expected tactile contact—such as which finger, which part, and how much pressure is needed—serving as physical world goals.
Reactive Error Correction: Upon touching an object, the robot continuously reads tactile and joint states, making minor corrections to the original motion to avoid failures where it "looks right but actually fails."
Three-Tier Time Scales: The system is split into a 1Hz high-level task planner with visual-tactile goal prediction, a 10Hz mid-level motion generator, and a 30Hz tactile reactive layer for rapid correction. This prevents tactile data from being overwhelmed by vision if forced into a unified VLA.
Experimental Results: Across six real-robot tasks (watering plants, desk clearing, plugging in power, inserting cups, scrubbing pots, pulling tissues), the full system achieved an average success rate of 65.0% / 57.2% in undisturbed/disturbed scenarios. Performance dropped significantly when removing tactile input, the 30Hz correction layer, the tactile world model, or the sub-task planning module.
Data Pipeline: The team previously developed EgoTouch and TouchAnything to solve tactile data collection and tactile recovery from multi-view videos, respectively. TouchWorld closes this loop by applying the pipeline to actual robot control.
Engineering & Plans: The team is building custom hardware, data pipelines, and evaluation systems. They plan to progressively open-source relevant data, code, and models, while continuing to advance five-finger dexterous hands and whole-body mobile manipulation.

Original post →

Tactile World Model Enables Robot Error Correction

More from Embodied